General Metropolis-Hastings jump diffusions for automatic target recognition in infrared scenes

To locate and recognize ground-based targets in forward- looking IR (FLIR) images, 3-D faceted models with associated pose pa- rameters are formulated to accommodate the variability found in FLIR imagery. Taking a Bayesian approach, scenes are simulated from the emissive characteristics of the CAD models and compared with the col- lected data by a likelihood function based on sensor statistics. This like- lihood is combined with a prior distribution defined over the set of pos- sible scenes to form a posterior distribution. To accommodate scenes with variable numbers of targets, the posterior distribution is defined over parameter vectors of varying dimension. An inference algorithm based on Metropolis-Hastings jump-diffusion processes empirically samples from the posterior distribution, generating configurations of templates and transformations that match the collected sensor data with high prob- ability. The jumps accommodate the addition and deletion of targets and the estimation of target identities; diffusions refine the hypotheses by drifting along the gradient of the posterior distribution with respect to the orientation and position parameters. Previous results on jumps strate- gies analogous to the Metropolis acceptance/rejection algorithm, with proposals drawn from the prior and accepted based on the likelihood, are extended to encompass general Metropolis-Hastings proposal den- sities. In particular, the algorithm proposes moves by drawing from the posterior distribution over computationally tractible subsets of the param- eter space. The algorithm is illustrated by an implementation on a Silicon Graphics Onyx/Reality Engine.